Human detection and localization using aerial infrared images for post disaster management

Abstract

During search and rescue missions, the rescue team must look for potential victims after a disaster has taken place and bring them to safety as quickly as possible. Lack of visibility of the victims and their location can waste time and resources and significantly hinder this procedure. In this project, we explore the theory behind deep learning, computer vision specifically object detection, and develop deep learning methods for object detection from aerial infrared (FLIR) thermal images taken by a UAV, with applications in the development of navigational aids for search and rescue operations. We demonstrate that a Faster R-CNN object detection network can be trained and fine-tuned to detect specific objects (survivors) in aerial images. We also seek to demonstrate that using multiple datasets in combination during the training of the network can provide significant performance improvements on test data.